Deep learning based normative models identify structures that become abnormal with Alzheimer’s disease

Abstract

AbstractBackgroundImproving dementia diagnosis is a major global concern, given that up to 20% of subjects are misdiagnosed. Normative modelling, which quantifies how far individuals deviate from healthy controls (HC), may help to better situate subjects along the spectrum. To determine whether this is the case, we implemented a deep learning based normative model and tested it on structural neuroimaging data from subjects with HC, mild cognitive impairment (MCI), and Alzheimer’s disease (AD).MethodWe developed an adversarial autoencoder (AAE) (Figure 1) leveraging FreeSurfer features, namely cortical surface area of 68 cortical subregions (34 by hemisphere; Desikan‐Killiany atlas) and volumes of 33 neuroanatomical structures (Aseg atlas). Our model not only returns individual “deviation” scores—divergence from HC—but also identifies regions of interest that contribute to said score, facilitating interpretability. We retrieved T1‐weighted MRI scans (n = 204) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (HC/MCI/AD 105/29/60; Table 1). We trained our AAE on 71 HC samples (n = 71) and tested it on the rest (n = 133). We used bootstrapping to compute confidence intervals (CI) for each point estimate and examined regions where scores differed for the considered groups [1].ResultDeviation scores were lowest for HC (mean 0.74 [95%‐CI 0.63–0.88]), highest for AD (mean 1.21 [95%‐CI 1.03–1.50]), and intermediate for MCI (mean 1.04 [95%‐CI 0.88–1.27]), implying subjects indeed fell along the AD spectrum (Figure 2A). Said scores also enabled classifying subjects into HC, MCI, and AD with a mean area under the curve (AUC) of 75% (Figure 2B‐C; HC vs. MCI: 0.75 [95%‐CI 0.62–0.85]; HC vs. AD: 0.76 [95%‐CI 0.64–0.82]). According to the AAE, the hippocampus, parahippocampal cortex, ventricles, and inferior parietal lobe contributed to such differences (Figure 2D‐E).ConclusionNormative models can identify structures that become abnormal as the brain undergoes ageing and AD, in line with the literature. Further testing on a larger and more heterogeneous sample will reveal the potential of this tool and its potential deployment in clinics.